Skip to main content

​
Evidence Discipline

Most AI failures are not spectacular hallucinations. More often, a plausible intermediate result gets promoted into a stronger conclusion:
  • a command exits successfully, so the agent says the system is fixed;
  • a file was changed, so the agent says the feature works;
  • a local build passes, so the agent says the deployment is live;
  • a search returns no result, so the agent says nothing exists;
  • you state a premise, so the agent repeats it as independently verified fact.
Myrm’s Evidence Discipline skill makes these failures rare by enforcing a claim-to-evidence contract on every answer the default agent (builtin-economy) produces.

​
What it does

The skill separates claims by evidence type and applies a proof obligation scaled to how strong the claim is:

​
Safety contract

Beyond labeling evidence, the skill blocks the common ways agents quietly lie:
  • Never claim access to resources that were not actually available in the task.
  • Never present a plan or an attempted action as an action that already happened.
  • Never reinterpret tool errors, empty results, partial sync, or inaccessible data as successful verification.
  • Never perform unrelated destructive actions merely to gain stronger evidence.
  • Bound negative claims (e.g. “nothing exists”) to the surfaces actually checked.

​
Built-in balance against over-caution

Evidence discipline cuts both ways, so the skill also prevents the opposite failure:
  • Don’t overuse “UNKNOWN”. When evidence is reasonably available, verify it—don’t use caution as an excuse to skip the check.
  • Check before you ask. If you can inspect the evidence yourself, do it before asking the user to repeat information.
The result: an agent that is honest about what it doesn’t know, actively verifies what it can, and never offloads its own homework onto you.

​
How to use it

Zero configuration. The skill is bound by default to the built-in general agent (builtin-economy) and active on every conversation—including Web, scheduled tasks, and channels. You can also:
  • View it under Settings → Skills — it appears alongside other prebuilt skills.
  • Bind it to any agent you create, or remove it from an agent that doesn’t need it.
  • Give it a stronger trigger in an agent’s custom system prompt, e.g. “Prove important claims before you answer.”

​
Verification & contract tests

The skill’s behavioral contract is guarded by 33 architecture contract tests that assert every clause is present and the frontmatter version stays valid, so the skill can’t silently drift or degrade. The full pipeline (sync → binding → visibility) is covered by integration tests without mocks.